Meituan has open-sourced LongCat-2.5 under an MIT license, permitting commercial
inference deployment and further training without additional authorization.
LongCat-2.5 is a native multimodal model for long-range agent tasks, with
improved end-to-end performance in complex codebase development, full project
builds, visual coding and professional office workflows. The model totals 1.6
tln parameters, with per-token activation of ~48 bln parameters, and supports a
one‑million‑token context. Meituan says the model architecture, chip adaptation
and deployment strategies were co-optimized for domestic compute chips with
constrained VRAM and bandwidth.
